# Chapter 3 · A command is not a sound · brief · 11 September 2026

The third chapter of the _Tara · practise_ track, rewritten in the Eve edition
on 11 September 2026. It follows the plan in `EVE-V1-REDESIGN-TODOS.md` and the
track brief in [tara-practise-track.md](tara-practise-track.md). The receipt is
under `verification/`.

## The job

Chapter two ended on an audio floor of eighty per cent that no template lowers.
This chapter is where that percentage — and the position, the progress and the
listening time beside it — actually comes from.

The shared audio model keeps two phases, not one. A session phase says where the
practice stands in eight steps; a playback phase says what the transport is
reported to be doing in eight more. They move independently: loading tracks
leaves a session ready while playback is still idle, and an interruption leaves
the session paused while playback reads interrupted.

Two paths write to that record and they do not share rules. Commands go through
a transition table of eighty phase-and-event combinations and refuse anything
not in it. Engine updates bypass that table entirely: they can revise a
completed session, emit a completion and an error from one update, and copy a
raw position the seek path would have clamped.

And the one number that sounds like a measurement is the narrowest of all.
`TaraPlaybackEngineUpdate` has eleven fields and none of them is a clock, so
`activeListenSeconds` adds the host's own wall-clock time between two _calls_ —
`Date.now()` at this call minus the reading the previous call stored — and only
when the previous playback phase was `playing` and the incoming report says
playing. It never reads the media position, never multiplies by the rate, and
never inspects whether anything was audible. Because the interval is the calling
cadence, an executed run with reports two seconds apart accumulates four seconds
of listening while the track advances two.

**Running example.** One 600-second practice. A call half a second after the
last one jumps the position to 250 and adds half a second of listening; a
buffering call a second later that still says playing adds one more; the
recovery call after it adds nothing. The practice is interrupted, the caller
decides whether ending the interruption resumes it, and the session finishes at
120 seconds with progress 0.2 and 1.5 seconds listened. Every value here is
executed against the pinned source under a controlled host clock.

## Three learning promises (the cover)

1. **Two phases, moving independently**, and a screen that reads only one of
   them will be wrong about the other.
2. **Commands and engine updates are separate paths**, and only one of them
   respects the transition table.
3. **Listening time is wall clock between qualifying calls** — not position, not
   rate, not attention.

## Three things to remember (the close)

1. **An accepted command is not an observed sound.** The manager assigns state
   and emits an event before any adapter has called a player.
2. **Terminal is a command boundary, not a guarantee.** An engine update can
   move a completed session to error.
3. **Nothing here saves anything.** A snapshot is a value for a consumer that
   has to exist and does not yet.

**Bridge.** Chapter four — Tara ritual assembly and tone review: the moments a
practice is assembled from before any of this runs.

## Sections

Four sections of six, five, six and six.

| Section                          | Slides                                                                                        |
| -------------------------------- | --------------------------------------------------------------------------------------------- |
| Two records, and what each says  | the case, the two phases, the handoff, identity, the media, the lifecycle                     |
| What a command may do            | the live phases, the ending phases, seek, the interruption branch, the feedback loop          |
| What a report does to the record | the priority order, position, listening time, the two completions, a late update, the markers |
| Sampling, listeners and the end  | marker sampling, subscriptions, snapshots, a throwing listener, replacement, the closing case |

## Treatment decisions

- **Three recovery maps, one state machine.** The plan turned all three into
  state machines. `tara-audio-start` became the whole command lifecycle in one
  diagram — preparing, ready, active, paused and a grouped terminal box — which
  leaves nothing for a second machine to draw. `tara-audio-interrupt` is the
  `shouldResume` branch, so it is a decision tree; `tara-audio-replace-destroy`
  is two operations against the same terminal outcome, so it is a card grid.
- **`tara-audio-snapshots` became a compare panel.** Its rows already asked one
  question twice, of a shallow session copy against a scalar snapshot.
- **`tara-audio-status-priority` became a layer stack of five bands.** It was
  planned as a card grid grown from three cards to five, one per playback label,
  because the order is the content and three cards had to group two of the five
  together. Five cards wrap to two rows and ran 97px past the 1280×720 budget;
  more to the point, a cascade is a precedence stack, and each band's `boundary`
  states exactly the condition that sends a report to the band below. The
  library already uses `layer-stack` for precedence in `eve-policy-order`.
- **Every case study carries three beats.** Four fit at 1440 and run 6px past
  the narration chrome at 1280 — the Eve-edition `.case-evidence` list stacks
  rather than wrapping, so the fourth beat costs a full 88px row. The three
  chapter case studies that wanted a fourth (`tara-audio-listening`,
  `tara-audio-marker-sampling`, `tara-audio-recovery`) keep the fourth sample in
  their notes instead, where it reads as the extension it is.
- **No table.** The eighty-combination command matrix is split across two card
  grids of four phases, as the plan asks. Chapter two's transition table already
  carries the library's one state-table-as-table, and a second adjacent one
  would teach the form rather than the content.

## What this chapter does not establish

**No capture, and this is the one chapter in the track that has none.** The
shared audio model's only consumer is native, no pinned capture of that surface
exists, and the web player from chapter one does not use this manager at all —
so a capture here would have to show a screen the chapter is not about. The plan
item that asks for one is left unchecked rather than satisfied by a picture that
would mislead.

Nothing else, either: no persistence, no device, no audibility. The manager
writes no store; a snapshot is eleven scalars handed to a consumer that has to
exist. No native acceptance — the adapter is read in source and the linked
native guide follows the actual mobile caller. And `activeListenSeconds` is a
source measurement of qualifying wall-clock intervals, not a claim about
attention, benefit or completion quality.
